A Phenotyping of Diastolic Function by Machine Learning Improves Prediction of Clinical Outcomes in Heart Failure

Abstract

Background: Discriminating between different patterns of diastolic dysfunction in heart failure (HF) is still challenging. We tested the hypothesis that an unsupervised machine learning algorithm would detect heterogeneity in diastolic function and improve risk stratification compared with recommended consensus criteria. Methods: This study included 279 consecutive patients aged 24-97 years old with clinically stable HF referred for echocardiographic assessment, in whom diastolic variables were measured according to the current guidelines. Cluster analysis was undertaken to identify homogeneous groups of patients with similar profiles of the variables. Sequential Cox models were used to compare cluster-based classification with guidelines-based classification for predicting clinical outcomes. The primary endpoint was hospitalization for worsening HF. Results: The analysis identified three clusters with distinct properties of diastolic function that shared similarities with guidelines-based classification. The clusters were associated with brain natriuretic peptide level (p < 0.001), hemoglobin concentration (p = 0.017) and estimated glomerular filtration rate (p = 0.001). During a mean follow-up period of 2.6 ± 2.0 years, 62 patients (22%) experienced the primary endpoint. Cluster-based classification predicted events with a hazard ratio 1.68 (p = 0.019) that was independent from and incremental to the Meta-analysis Global Group in Chronic Heart Failure (MAGGIC) risk score for HF, and from left ventricular end-diastolic volume and global longitudinal strain, whereas guidelines-based classification did not retain its independent prognostic value (hazard ratio = 1.25, p = 0.202). Conclusion: Machine learning can identify patterns of diastolic function that better stratify the risk for decompensation than the current consensus recommendations in HF. Integrating this data-driven phenotyping may help in refining prognostication and optimizing treatment.

Keywords: diastolic function; echocardiogram classification; heart failure; machine learning; prognostication factor.

Figure 1

Study design.

Figure 2

Comparisons of diastolic function variables. These five variables (A–E) were used for cluster analysis. The e' (B) decreased and E/e' (C), LAVi (D) and TRV (E) increased, as diastolic function worsened (grade/cluster number increased). LAVi, left atrium volume index; TRV, tricuspid regurgitation velocity.

Figure 3

Bayesian information criterion. This result demonstrated that three-cluster model fit the best, because the absolute value of Bayesian information criterion was the lowest when the dataset was modeled with three clusters.

Figure 4

Comparisons of BNP, eGFR and hemoglobin level across grades and clusters. (A) Comparisons of BNP, (B) eGFR, and (C) hemoglobin level across grades and clusters. BNP, brain natriuretic peptide; eGFR, estimate glomerular filtration rate.

Figure 5

Kaplan–Meier curves stratified by grades and clusters. (A) Primary endpoint (WHF) and (B) secondary endpoint (a composite of CV deaths and WHF) when stratified by guidelines-based classification. (C) Primary endpoint and (D) secondary endpoint when stratified by cluster-based classification. WHF, worsening heart failure; CV, cardiovascular.

Figure 6

Nested Cox models. A baseline Cox model was first constructed with MAGGIC score, LVEDV and LVGLS. A nested model was then constructed by adding grades and cluster separately. (A) For primary endpoint. (B) For secondary endpoint. MAGGIC, Meta-analysis Global Group in Chronic Heart Failure; LVEDV, left ventricular end-diastolic volume; LVGLS, left ventricular global longitudinal strain.

Figure 7

Comparison of grade and cluster distribution. Grade (A) and cluster (B) distribution in the first 2 dimensions identified by principal component analysis. (C) Mahalanobis distance from each subject to the center of cluster 1 was color-coded.

Figure 8

Clinical validations of clusters in the subgroup of all 188 patients with grade 1 diastolic dysfunction by echocardiographic criteria. (A) Comparison of BNP level, (B,C) clinical outcomes [(B) for primary endpoint, (C) for secondary endpoint] stratified by clusters. BNP, brain natriuretic peptide; WHF, worsening heart failure; CV, cardiovascular.